Added a method to reset the neural network to the best state when usi…

[leconteur]

…ng n_stable argument

[leconteur]

Did you check on this PR?

[alexjc]

I've tested this finally, I like the change!

I think we should just save the best network within _train_layer locally and restore it immediately once that function is finished. Is there a use case for manually resetting the network? Why do we want sub-optimal weights?

[leconteur]

I agree.

I only did the change in the pylearn2 backend however. I don't remember whether it is possible to implement this in the other backends. If that is not the case, maybe we should only enable it explicitly to remain coherent across the different backends.

[alexjc]

There is only one functional backend right now, so if you change it for pylearn2 that's fine for this PR. Also might be a good idea to document this somewhere...

[alexjc]

This is now implemented in 2694667. Thanks for looking into this, it helped a lot!

[leconteur]

I was glad to have helped.